3-2_MNIST_MLP_Dropout_ReLU.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "''' 1. Module Import '''\n",
    "import numpy as np\n",
    "import matplotlib.pyplot as plt\n",
    "\n",
    "import torch\n",
    "import torch.nn as nn\n",
    "import torch.nn.functional as F\n",
    "from torchvision import transforms, datasets"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Using PyTorch version: 1.6.0+cu101  Device: cuda\n"
     ]
    }
   ],
   "source": [
    "''' 2. 딥러닝 모델을 설계할 때 활용하는 장비 확인 '''\n",
    "if torch.cuda.is_available():\n",
    "    DEVICE = torch.device('cuda')\n",
    "else:\n",
    "    DEVICE = torch.device('cpu')\n",
    "print('Using PyTorch version:', torch.__version__, ' Device:', DEVICE)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "BATCH_SIZE = 32\n",
    "EPOCHS = 10"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "''' 3. MNIST 데이터 다운로드 (Train set, Test set 분리하기) '''\n",
    "train_dataset = datasets.MNIST(root = \"../data/MNIST\",\n",
    "                               train = True,\n",
    "                               download = True,\n",
    "                               transform = transforms.ToTensor())\n",
    "\n",
    "test_dataset = datasets.MNIST(root = \"../data/MNIST\",\n",
    "                              train = False,\n",
    "                              transform = transforms.ToTensor())\n",
    "\n",
    "train_loader = torch.utils.data.DataLoader(dataset = train_dataset,\n",
    "                                           batch_size = BATCH_SIZE,\n",
    "                                           shuffle = True)\n",
    "\n",
    "test_loader = torch.utils.data.DataLoader(dataset = test_dataset,\n",
    "                                          batch_size = BATCH_SIZE,\n",
    "                                          shuffle = False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "X_train: torch.Size([32, 1, 28, 28]) type: torch.FloatTensor\n",
      "y_train: torch.Size([32]) type: torch.LongTensor\n"
     ]
    }
   ],
   "source": [
    "''' 4. 데이터 확인하기 (1) '''\n",
    "for (X_train, y_train) in train_loader:\n",
    "    print('X_train:', X_train.size(), 'type:', X_train.type())\n",
    "    print('y_train:', y_train.size(), 'type:', y_train.type())\n",
    "    break"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 720x72 with 10 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "''' 5. 데이터 확인하기 (2) '''\n",
    "pltsize = 1\n",
    "plt.figure(figsize=(10 * pltsize, pltsize))\n",
    "for i in range(10):\n",
    "    plt.subplot(1, 10, i + 1)\n",
    "    plt.axis('off')\n",
    "    plt.imshow(X_train[i, :, :, :].numpy().reshape(28, 28), cmap = \"gray_r\")\n",
    "    plt.title('Class: ' + str(y_train[i].item()))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "''' 6. Multi Layer Perceptron (MLP) 모델 설계하기 '''\n",
    "class Net(nn.Module):\n",
    "    def __init__(self):\n",
    "        super(Net, self).__init__()\n",
    "        self.fc1 = nn.Linear(28 * 28, 512)\n",
    "        self.fc2 = nn.Linear(512, 256)\n",
    "        self.fc3 = nn.Linear(256, 10)\n",
    "        self.dropout_prob = 0.5\n",
    "\n",
    "    def forward(self, x):\n",
    "        x = x.view(-1, 28 * 28)\n",
    "        x = self.fc1(x)\n",
    "        x = F.relu(x)\n",
    "        x = F.dropout(x, training = self.training, p = self.dropout_prob)\n",
    "        x = self.fc2(x)\n",
    "        x = F.relu(x)\n",
    "        x = F.dropout(x, training = self.training, p = self.dropout_prob)\n",
    "        x = self.fc3(x)\n",
    "        x = F.log_softmax(x, dim = 1)\n",
    "        return x"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Net(\n",
      "  (fc1): Linear(in_features=784, out_features=512, bias=True)\n",
      "  (fc2): Linear(in_features=512, out_features=256, bias=True)\n",
      "  (fc3): Linear(in_features=256, out_features=10, bias=True)\n",
      ")\n"
     ]
    }
   ],
   "source": [
    "''' 7. Optimizer, Objective Function 설정하기 '''\n",
    "model = Net().to(DEVICE)\n",
    "optimizer = torch.optim.SGD(model.parameters(), lr = 0.01, momentum = 0.5)\n",
    "criterion = nn.CrossEntropyLoss()\n",
    "\n",
    "print(model)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "''' 8. MLP 모델 학습을 진행하며 학습 데이터에 대한 모델 성능을 확인하는 함수 정의 '''\n",
    "def train(model, train_loader, optimizer, log_interval):\n",
    "    model.train()\n",
    "    for batch_idx, (image, label) in enumerate(train_loader):\n",
    "        image = image.to(DEVICE)\n",
    "        label = label.to(DEVICE)\n",
    "        optimizer.zero_grad()\n",
    "        output = model(image)\n",
    "        loss = criterion(output, label)\n",
    "        loss.backward()\n",
    "        optimizer.step()\n",
    "\n",
    "        if batch_idx % log_interval == 0:\n",
    "            print(\"Train Epoch: {} [{}/{} ({:.0f}%)]\\tTrain Loss: {:.6f}\".format(\n",
    "                epoch, batch_idx * len(image), \n",
    "                len(train_loader.dataset), 100. * batch_idx / len(train_loader), \n",
    "                loss.item()))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "''' 9. 학습되는 과정 속에서 검증 데이터에 대한 모델 성능을 확인하는 함수 정의 '''\n",
    "def evaluate(model, test_loader):\n",
    "    model.eval()\n",
    "    test_loss = 0\n",
    "    correct = 0\n",
    "\n",
    "    with torch.no_grad():\n",
    "        for image, label in test_loader:\n",
    "            image = image.to(DEVICE)\n",
    "            label = label.to(DEVICE)\n",
    "            output = model(image)\n",
    "            test_loss += criterion(output, label).item()\n",
    "            prediction = output.max(1, keepdim = True)[1]\n",
    "            correct += prediction.eq(label.view_as(prediction)).sum().item()\n",
    "    \n",
    "    test_loss /= (len(test_loader.dataset) / BATCH_SIZE)\n",
    "    test_accuracy = 100. * correct / len(test_loader.dataset)\n",
    "    return test_loss, test_accuracy"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train Epoch: 1 [0/60000 (0%)]\tTrain Loss: 2.293252\n",
      "Train Epoch: 1 [6400/60000 (11%)]\tTrain Loss: 2.103498\n",
      "Train Epoch: 1 [12800/60000 (21%)]\tTrain Loss: 1.014423\n",
      "Train Epoch: 1 [19200/60000 (32%)]\tTrain Loss: 0.739293\n",
      "Train Epoch: 1 [25600/60000 (43%)]\tTrain Loss: 0.824886\n",
      "Train Epoch: 1 [32000/60000 (53%)]\tTrain Loss: 0.373551\n",
      "Train Epoch: 1 [38400/60000 (64%)]\tTrain Loss: 0.590951\n",
      "Train Epoch: 1 [44800/60000 (75%)]\tTrain Loss: 0.609839\n",
      "Train Epoch: 1 [51200/60000 (85%)]\tTrain Loss: 0.725992\n",
      "Train Epoch: 1 [57600/60000 (96%)]\tTrain Loss: 0.316456\n",
      "\n",
      "[EPOCH: 1], \tTest Loss: 0.0099, \tTest Accuracy: 91.00 % \n",
      "\n",
      "Train Epoch: 2 [0/60000 (0%)]\tTrain Loss: 0.434049\n",
      "Train Epoch: 2 [6400/60000 (11%)]\tTrain Loss: 0.423664\n",
      "Train Epoch: 2 [12800/60000 (21%)]\tTrain Loss: 0.223913\n",
      "Train Epoch: 2 [19200/60000 (32%)]\tTrain Loss: 0.185579\n",
      "Train Epoch: 2 [25600/60000 (43%)]\tTrain Loss: 0.516311\n",
      "Train Epoch: 2 [32000/60000 (53%)]\tTrain Loss: 0.170941\n",
      "Train Epoch: 2 [38400/60000 (64%)]\tTrain Loss: 0.679857\n",
      "Train Epoch: 2 [44800/60000 (75%)]\tTrain Loss: 0.206488\n",
      "Train Epoch: 2 [51200/60000 (85%)]\tTrain Loss: 0.293542\n",
      "Train Epoch: 2 [57600/60000 (96%)]\tTrain Loss: 0.532575\n",
      "\n",
      "[EPOCH: 2], \tTest Loss: 0.0070, \tTest Accuracy: 93.47 % \n",
      "\n",
      "Train Epoch: 3 [0/60000 (0%)]\tTrain Loss: 0.494590\n",
      "Train Epoch: 3 [6400/60000 (11%)]\tTrain Loss: 0.692674\n",
      "Train Epoch: 3 [12800/60000 (21%)]\tTrain Loss: 0.141787\n",
      "Train Epoch: 3 [19200/60000 (32%)]\tTrain Loss: 0.254909\n",
      "Train Epoch: 3 [25600/60000 (43%)]\tTrain Loss: 0.143211\n",
      "Train Epoch: 3 [32000/60000 (53%)]\tTrain Loss: 0.290160\n",
      "Train Epoch: 3 [38400/60000 (64%)]\tTrain Loss: 0.370776\n",
      "Train Epoch: 3 [44800/60000 (75%)]\tTrain Loss: 0.711377\n",
      "Train Epoch: 3 [51200/60000 (85%)]\tTrain Loss: 0.091218\n",
      "Train Epoch: 3 [57600/60000 (96%)]\tTrain Loss: 0.176025\n",
      "\n",
      "[EPOCH: 3], \tTest Loss: 0.0054, \tTest Accuracy: 94.96 % \n",
      "\n",
      "Train Epoch: 4 [0/60000 (0%)]\tTrain Loss: 0.298294\n",
      "Train Epoch: 4 [6400/60000 (11%)]\tTrain Loss: 0.155041\n",
      "Train Epoch: 4 [12800/60000 (21%)]\tTrain Loss: 0.358912\n",
      "Train Epoch: 4 [19200/60000 (32%)]\tTrain Loss: 0.178691\n",
      "Train Epoch: 4 [25600/60000 (43%)]\tTrain Loss: 0.085475\n",
      "Train Epoch: 4 [32000/60000 (53%)]\tTrain Loss: 0.272500\n",
      "Train Epoch: 4 [38400/60000 (64%)]\tTrain Loss: 0.203928\n",
      "Train Epoch: 4 [44800/60000 (75%)]\tTrain Loss: 0.455426\n",
      "Train Epoch: 4 [51200/60000 (85%)]\tTrain Loss: 0.123246\n",
      "Train Epoch: 4 [57600/60000 (96%)]\tTrain Loss: 0.114354\n",
      "\n",
      "[EPOCH: 4], \tTest Loss: 0.0044, \tTest Accuracy: 95.63 % \n",
      "\n",
      "Train Epoch: 5 [0/60000 (0%)]\tTrain Loss: 0.124653\n",
      "Train Epoch: 5 [6400/60000 (11%)]\tTrain Loss: 0.157264\n",
      "Train Epoch: 5 [12800/60000 (21%)]\tTrain Loss: 0.202116\n",
      "Train Epoch: 5 [19200/60000 (32%)]\tTrain Loss: 0.063491\n",
      "Train Epoch: 5 [25600/60000 (43%)]\tTrain Loss: 0.215543\n",
      "Train Epoch: 5 [32000/60000 (53%)]\tTrain Loss: 0.039489\n",
      "Train Epoch: 5 [38400/60000 (64%)]\tTrain Loss: 0.062817\n",
      "Train Epoch: 5 [44800/60000 (75%)]\tTrain Loss: 0.074638\n",
      "Train Epoch: 5 [51200/60000 (85%)]\tTrain Loss: 0.100656\n",
      "Train Epoch: 5 [57600/60000 (96%)]\tTrain Loss: 0.263398\n",
      "\n",
      "[EPOCH: 5], \tTest Loss: 0.0038, \tTest Accuracy: 96.28 % \n",
      "\n",
      "Train Epoch: 6 [0/60000 (0%)]\tTrain Loss: 0.243492\n",
      "Train Epoch: 6 [6400/60000 (11%)]\tTrain Loss: 0.134330\n",
      "Train Epoch: 6 [12800/60000 (21%)]\tTrain Loss: 0.132107\n",
      "Train Epoch: 6 [19200/60000 (32%)]\tTrain Loss: 0.210380\n",
      "Train Epoch: 6 [25600/60000 (43%)]\tTrain Loss: 0.166309\n",
      "Train Epoch: 6 [32000/60000 (53%)]\tTrain Loss: 0.154378\n",
      "Train Epoch: 6 [38400/60000 (64%)]\tTrain Loss: 0.415324\n",
      "Train Epoch: 6 [44800/60000 (75%)]\tTrain Loss: 0.091827\n",
      "Train Epoch: 6 [51200/60000 (85%)]\tTrain Loss: 0.088081\n",
      "Train Epoch: 6 [57600/60000 (96%)]\tTrain Loss: 0.156756\n",
      "\n",
      "[EPOCH: 6], \tTest Loss: 0.0035, \tTest Accuracy: 96.58 % \n",
      "\n",
      "Train Epoch: 7 [0/60000 (0%)]\tTrain Loss: 0.311500\n",
      "Train Epoch: 7 [6400/60000 (11%)]\tTrain Loss: 0.153611\n",
      "Train Epoch: 7 [12800/60000 (21%)]\tTrain Loss: 0.182327\n",
      "Train Epoch: 7 [19200/60000 (32%)]\tTrain Loss: 0.109722\n",
      "Train Epoch: 7 [25600/60000 (43%)]\tTrain Loss: 0.357384\n",
      "Train Epoch: 7 [32000/60000 (53%)]\tTrain Loss: 0.124124\n",
      "Train Epoch: 7 [38400/60000 (64%)]\tTrain Loss: 0.139059\n",
      "Train Epoch: 7 [44800/60000 (75%)]\tTrain Loss: 0.094896\n",
      "Train Epoch: 7 [51200/60000 (85%)]\tTrain Loss: 0.069713\n",
      "Train Epoch: 7 [57600/60000 (96%)]\tTrain Loss: 0.060162\n",
      "\n",
      "[EPOCH: 7], \tTest Loss: 0.0031, \tTest Accuracy: 97.02 % \n",
      "\n",
      "Train Epoch: 8 [0/60000 (0%)]\tTrain Loss: 0.219568\n",
      "Train Epoch: 8 [6400/60000 (11%)]\tTrain Loss: 0.265314\n",
      "Train Epoch: 8 [12800/60000 (21%)]\tTrain Loss: 0.136442\n",
      "Train Epoch: 8 [19200/60000 (32%)]\tTrain Loss: 0.182477\n",
      "Train Epoch: 8 [25600/60000 (43%)]\tTrain Loss: 0.057788\n",
      "Train Epoch: 8 [32000/60000 (53%)]\tTrain Loss: 0.180467\n",
      "Train Epoch: 8 [38400/60000 (64%)]\tTrain Loss: 0.163830\n",
      "Train Epoch: 8 [44800/60000 (75%)]\tTrain Loss: 0.146127\n",
      "Train Epoch: 8 [51200/60000 (85%)]\tTrain Loss: 0.078550\n",
      "Train Epoch: 8 [57600/60000 (96%)]\tTrain Loss: 0.366327\n",
      "\n",
      "[EPOCH: 8], \tTest Loss: 0.0028, \tTest Accuracy: 97.16 % \n",
      "\n",
      "Train Epoch: 9 [0/60000 (0%)]\tTrain Loss: 0.032364\n",
      "Train Epoch: 9 [6400/60000 (11%)]\tTrain Loss: 0.049485\n",
      "Train Epoch: 9 [12800/60000 (21%)]\tTrain Loss: 0.083032\n",
      "Train Epoch: 9 [19200/60000 (32%)]\tTrain Loss: 0.120306\n",
      "Train Epoch: 9 [25600/60000 (43%)]\tTrain Loss: 0.236849\n",
      "Train Epoch: 9 [32000/60000 (53%)]\tTrain Loss: 0.034962\n",
      "Train Epoch: 9 [38400/60000 (64%)]\tTrain Loss: 0.355975\n",
      "Train Epoch: 9 [44800/60000 (75%)]\tTrain Loss: 0.098553\n",
      "Train Epoch: 9 [51200/60000 (85%)]\tTrain Loss: 0.739664\n",
      "Train Epoch: 9 [57600/60000 (96%)]\tTrain Loss: 0.029625\n",
      "\n",
      "[EPOCH: 9], \tTest Loss: 0.0027, \tTest Accuracy: 97.21 % \n",
      "\n",
      "Train Epoch: 10 [0/60000 (0%)]\tTrain Loss: 0.316167\n",
      "Train Epoch: 10 [6400/60000 (11%)]\tTrain Loss: 0.184830\n",
      "Train Epoch: 10 [12800/60000 (21%)]\tTrain Loss: 0.160652\n",
      "Train Epoch: 10 [19200/60000 (32%)]\tTrain Loss: 0.162540\n",
      "Train Epoch: 10 [25600/60000 (43%)]\tTrain Loss: 0.044575\n",
      "Train Epoch: 10 [32000/60000 (53%)]\tTrain Loss: 0.090206\n",
      "Train Epoch: 10 [38400/60000 (64%)]\tTrain Loss: 0.021095\n",
      "Train Epoch: 10 [44800/60000 (75%)]\tTrain Loss: 0.275242\n",
      "Train Epoch: 10 [51200/60000 (85%)]\tTrain Loss: 0.117553\n",
      "Train Epoch: 10 [57600/60000 (96%)]\tTrain Loss: 0.028020\n",
      "\n",
      "[EPOCH: 10], \tTest Loss: 0.0026, \tTest Accuracy: 97.40 % \n",
      "\n"
     ]
    }
   ],
   "source": [
    "''' 10. MLP 학습 실행하며 Train, Test set의 Loss 및 Test set Accuracy 확인하기 '''\n",
    "for epoch in range(1, EPOCHS + 1):\n",
    "    train(model, train_loader, optimizer, log_interval = 200)\n",
    "    test_loss, test_accuracy = evaluate(model, test_loader)\n",
    "    print(\"\\n[EPOCH: {}], \\tTest Loss: {:.4f}, \\tTest Accuracy: {:.2f} % \\n\".format(\n",
    "        epoch, test_loss, test_accuracy))"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.8"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}